Function report_key 

pub fn report_key(
    head_sha: &str,
    changes: &[WorktreeChange],
    opts: &Options,
) -> Result<String>

The content key for a change set: lowercase-hex SHA-256 of head_sha | sorted "path\0content-sha256" lines | crate version | calib=<digest> | rows_limit=<n> | opts=<canonical-json digest> | [KEY_SCHEMA]. A deleted path contributes the literal "deleted" in place of its content hash.

The defect-calibration artifact’s CONTENT digest is folded in because --defect-calibration substitutes smell weights inside the scoring engine — it changes the MEASURED health scores in the report, not just the verdict. Two runs on the same worktree and HEAD, one calibrated and one not, must not share a cache entry. calib= is empty when no artifact is configured. Thresholds stay excluded: they only affect verdicts, which consumers always recompute from the cached report.

Every other report-affecting Options knob is folded in via Options::canonical_json — the same digest the ingest cache uses — so min_revs, exclude_patterns/include_ignored, the clone thresholds, and any future field are covered with zero per-field maintenance: build_change_set_report’s run_coupling(db, opts) and clone_introductions(db, opts, …) calls pass the caller’s opts straight through (unlike the health projection’s own opts_scan, which pins min_revs = 1), so any of those knobs can change which coupling-absence or clone-introduction findings this report contains. rows_limit is folded in explicitly ALONGSIDE the canonical digest rather than relying on it: canonical_json deliberately drops rows_limit as cosmetic for the ingest cache, but run_coupling(db, opts) truncates to it before the coupling-absence filter runs, so it is not cosmetic here.

Errors

Returns CodeLoreError::Analysis when a non-deleted change-set file cannot be read from the working tree — the engine could not project it either, so failing early is honest.